"Every time I read of a new algorithm for producing a hit pop song, or software for writing a Man Booker-worthy novel, I know I’m in for a disappointment"
Every time I read of a new algorithm for producing a hit pop song, or software for writing a Man Booker-worthy novel, I know I’m in for a disappointment. The latest is a computer program that generates sonnets. Sonnets! Fancy! Computer scientists at IBM and the universities of Melbourne and Toronto have built a system that creates poems so metrically accomplished that the majority of readers in a study failed to distinguish them from the efforts of humans.
The project is romantic in its way. We’ve been trying to build poetry machines for hundreds of years. In the mid-19th century, a British inventor constructed a particularly clever device which turned out Latin hexameters. John Clark exhibited his “Eureka” box in the Egyptian Hall on Piccadilly in 1845 and used it to conjure original verse as if by magic. How efficiently the Eureka completed what a schoolboy would struggle over for hours with his Latin primer.
It isn’t hard to see how a Eurekan hexameter could fool even a keen Latinist — but computer-generated English? The examples I’ve seen are barely comprehensible. It says something of how baffling poetry has become when a string of words set to metre passes for a sonnet.
It’s true that the language of poetry can be unpredictable. It was the strange placement of some of the epithets in Homer, after all, that helped scholars to reveal the oral tradition behind his epics. That Achilles could be described as “swift-footed” even when he was stationary hinted at the metrical usefulness of set phrases to an improvising bard. Scholars have moved away from seeing epithets as simply functional but, to the modern eye, such features remain peculiar. The obscurity of poetic imagery likewise.
The authors of the recent study had to concede that computer sonnets “still underperform in terms of readability and emotion” after they failed to convince a professor of English. The problem wasn’t simply the inferior quality of their language but the faultlessness of their metre. The computer system, to which the scientists input thousands of existing sonnets, achieved a higher degree of metrical accuracy and rhyme than was found in poems written by humans. The computer, in other words, failed to break the rules in the ways that real poets do.
It could not be taught that the best poets — Shakespeare, Donne, Pound, to name a few — have always played fast and loose with metrical schemes. Just think of Titus Andronicus’s lines when he’s at his most disturbed. Pope has often been criticised for his “corrections” to Shakespeare’s errant iambic pentameter. It takes a good eye to recognise when a poet has disrupted the metre intentionally. It takes a human to understand the reasons he might have done so.
A computer must learn to fail better if it is to produce anything truly poetic. Feed it poetry which shuns all pretensions to conformity. Some Edward Lear, perhaps, or Dr Seuss — it will learn that “nimbly” can rhyme with “chimbley”. Even some Auden would do. A poem containing “llamas” and “pyjamas” after O Tell Me The Truth About Love would be brilliant. Better still, do something revolutionary: teach the art of poetry to humans so they can write sonnets for themselves.